# AYORAI ATTRACTOR

> AYORAI ATTRACTOR is a Brazilian open-source project for evidence-based technical auditing of artificial intelligence answers. It verifies AI answers at the level of each individual claim: it relates each claim to evidence and provenance, analyzes whether the evidence supports, contradicts or does not clarify the claim, and a deterministic Judge with versioned rules issues the final verdict. Author, maintainer and technical director: Anderson Leon Ayora. License: Apache-2.0.

O AYORAI ATTRACTOR é um projeto brasileiro open source de auditoria técnica de respostas de inteligência artificial baseada em evidências, de autoria de Anderson Leon Ayora. Lema: "Da afirmação à evidência. Da evidência a um resultado verificável."

Core principle / Princípio fundamental: Intelligence ≠ Authorization ≠ Execution (Inteligência ≠ Autorização ≠ Execução). Models may assist the analysis but are not the authority for the final verdict; the deterministic Judge applies the system's explicit contract.

Judge 2.0 verdicts: SUPPORTED, PARTIALLY_SUPPORTED, UNVERIFIED, REFUTED, CONFLICTING.

Current version: 0.7.0. Version DOI: https://doi.org/10.5281/zenodo.24045951. Concept DOI: https://doi.org/10.5281/zenodo.23125082. Python 3.11, 3.12, 3.13.

Status and limits: research and engineering project. Not presented as certified; does not guarantee detection of every hallucination; does not by itself establish regulatory compliance. There is no Portuguese-language benchmark result yet. Calibration V2 on the English LLM-AggreFact dataset (30,420 DEV examples, 11 cohorts) selected FactCG with 74.68% cross-fitted balanced accuracy in DEV-A (MiniCheck 72.87%; mDeBERTa 62.57%). FactCG DEV-A metrics: AUROC 0.8650, AUPRC 0.9454, Brier 0.1168, ECE 0.0054. The official v0.7.0 LTT result remains 0/11 under the originally registered rule. A post-hoc exploratory re-aggregation under Amendment 2 found four DEV cohorts passing the amended procedure (RAGTruth, ClaimVerify, Lfqa, Reveal); this is not confirmatory certification, and seven cohorts did not reject the null at their first tested threshold. A proposed 367-accepted-sample floor for v0.8 is not an adopted rule and would make some smaller cohorts untestable by design. These are development results, not TEST, production performance, ranking or evidence of Portuguese generalization. Historical Golden v0 and v0.1 results are synthetic, in English and development-oriented. Hashes verify artifact integrity, not content truth.

## Official links

- [Official website](https://ayorinha.github.io/ayorai-opensearch/): project overview (Portuguese)
- [Guia de auditoria de inteligência artificial (PT)](https://ayorinha.github.io/ayorai-opensearch/auditoria-de-inteligencia-artificial/): what AI auditing is, what to audit, step-by-step verification of AI answers, open-source tools
- [Alucinação de IA: como detectar e verificar respostas (PT)](https://ayorinha.github.io/ayorai-opensearch/alucinacao-de-ia/): what AI hallucination is, warning signs, step-by-step verification of ChatGPT and other AI answers
- [Official GitHub repository](https://github.com/Ayorinha/ayorai-opensearch): source code, documentation, ADRs, evaluation
- [Technical status](https://github.com/Ayorinha/ayorai-opensearch/blob/main/docs/STATUS.md): current deliveries, checks and limitations
- [Public technical document v1.1 (PDF, PT)](https://github.com/Ayorinha/ayorai-opensearch/blob/main/docs/institutional/AYORAI-ATTRACTOR-Documento-Tecnico-v1.1.pdf)
- [Judge 2.0 contract and traceability](https://github.com/Ayorinha/ayorai-opensearch/blob/main/docs/judge2/TRACEABILITY.md)
- [Video presentation by the author (LinkedIn, 2026-11-05)](https://www.linkedin.com/feed/update/urn:li:activity:7524061945113493504/)
- [Zenodo DOI 10.5281/zenodo.23125082](https://doi.org/10.5281/zenodo.23125082): citable archive

## Author

- Anderson Leon Ayora — author, maintainer and technical director ([GitHub](https://github.com/Ayorinha))